Before close: the eight levers that act on price, structure and the first hundred days

The levers in this chapter do not move the portfolio company's P&L. They move the price paid, the structure of the deal, the risk the buyer retains, the cost of the diligence itself, and, in the last two, what the company is on day one hundred. That makes them the levers most often left out of an AI value creation plan, because the plan is usually written after the deal is priced. Written before, it changes the price.

Two survey numbers set the scene. KPMG's March 2026 survey of about 700 decision-makers at companies and private equity firms across twenty countries found 56% using AI in due diligence and valuation, 53% in sourcing and strategy, and 45% in post-merger integration.¹ Deloitte's 2025 survey of 1,000 senior investors found 86% adoption in M&A workflows and 35% using generative AI in due diligence specifically, and its 2026 update found 90% using it and fewer than a third with fully integrated tools.² Adoption is broad. Integration is not, and neither survey measures what the tools found.

The cost side is what makes the arithmetic interesting below $100M of enterprise value. ACG's July 2026 piece on the lower middle market puts a standard technology diligence engagement at $25,000 to $40,000, an expanded one at $40,000 to $75,000, a PE-grade one at up to $150,000, and a Big Four quality of earnings at $30,000 to $50,000 on a $5M acquisition.³ At $10M to $25M of enterprise value those fees are a prohibitive share of the deal, so the work does not get done, and "a networked former CTO parachuted in days before close with a two-week window and a checklist" is, in ACG's words, what passes for it. The same piece carries the caution that governs every card below: "AI-generated findings are fluent, and fluency is not accuracy."

The pre-close cards use a different metrics block from the operating levers, because there is no year one or year three: the columns are what the work typically costs and finds today, what it costs and finds with the lever, and which part of the model the output feeds.

P1 · Code and architecture diligence

What it is. An AI-assisted review of the target's codebase, dependency tree, test coverage, infrastructure and commit history, run on the whole repository rather than a sample, with the findings converted into a remediation estimate and a capex reserve in the model. It replaces, or sits underneath, the two-week expert review.

The line it moves. Price paid; the capex reserve and the R&D line in the model; the timing of O2 (migrations) and G4 (hosting).

Typical range and the evidence. Medium for the mechanism, weak for any vendor's specific claim. What has changed is that the commit history has become a diligence artifact in itself. GitClear's analysis of 623 million code changes between 2023 and 2026 found refactoring line moves down 70% and code block duplication up 81%, with copy-paste prevalence rising from 9.4% to 15.7%, all of it computable from a target's repository in hours.⁴ Veracode's July 2025 test of more than a hundred models found 45% of AI-generated code samples failed security tests and introduced OWASP Top 10 vulnerabilities, 72% in Java, and that "security performance remained flat, regardless of model size or training sophistication."⁵ Those two findings together mean a target that adopted coding agents in 2024 or 2025 may carry more debt than its release cadence suggests, and the debt is measurable. Commercial code-diligence tools claim an 80% reduction in manual CTO review time and name customers, but publish no methodology; treat those figures as a rate card, not a result.⁶

Conditions. Read access to the full repository and CI history under the NDA; a reviewer who can distinguish a real finding from a fluent one; a defined threshold for what becomes a price adjustment versus a hundred-day task.

Kill criteria. This lever is cheap enough that the kill criterion is quality rather than cost: if the automated review's top ten findings are not reproduced by a human reading the flagged files, the output is discarded and the traditional review is commissioned.

Horizon and stage. Deploy. Diligence.

How it translates. A: a twelve-year-old codebase with a modern front end, where the finding is the size of the test-coverage gap and the migration cost. B: a rules engine and content pipeline, where the finding is whether the regulatory logic is testable. C: a payments integration and a mobile app, where the finding is PCI scope and the processor dependency. D: a large modern codebase where the finding is the AI-generated share and its defect trend. E: multiple acquired codebases, where the finding is how many there are.

P2 · Contract and data-rights review

What it is. Every customer, vendor, partner and employment agreement read for assignment and change-of-control clauses, price escalators, renewal terms, data-use and data-ownership language, AI-use restrictions, business associate and data processing agreements, and indemnities, with the output structured into a database the company keeps after close. It decides what the revenue levers are legally allowed to do.

The line it moves. Price; representations and warranties; the feasibility of R2, R10 and X3, because a claimed data moat is only real if the customer contracts permit the use.

Typical range and the evidence. Medium. Harvey's July 2025 methodology note reports Bridgewater Associates cutting vendor contract review "from an average of 2 days to 2 hours" and Deutsche Telekom's in-house team reclaiming "up to 5 hours per lawyer per week," with the method (surveys, query telemetry, time-and-motion studies) disclosed; the aggregate "60 to 90% more efficient" is self-reported and should not be treated as measured.⁷ Hebbia's OpenAI-published case reports private equity firms saving "20 to 30 hours per deal," with no customer named.⁸ The counterweight is peer-reviewed: Stanford's benchmark of legal research tools found them producing incorrect information "more than 17% of the time" at best and more than 34% at worst, on open-ended research rather than clause extraction, which is a narrower and more checkable task, so those rates are an upper bound on the wrong job.⁹ The legacy contract-AI vendors have published nothing quantified since 2021.

The genuinely new finding is a gap in the record: no published figure exists for how often software targets lack the contractual right to use customer data for model training or product improvement. Whether that right exists is now the single most important line in the data room for the Invent levers, and it is the lever where the reviewer's judgment, not the tool's speed, sets the value.

Conditions. The full contract set in the data room, including amendments and order forms; a playbook of the clauses that matter and the positions the buyer will take on each; counsel who will stand behind the extraction.

Kill criteria. A 5% sample of extracted terms re-read by a person; an error rate above 2% on money terms (price, escalator, term, termination) sends the whole set back.

Horizon and stage. Deploy. Diligence.

How it translates. A: BAAs and state Medicaid terms decide what an agent may see. B: data-use rights across customers' regulatory submissions are the asset being bought. C: merchant agreements and processor contracts decide whether payments revenue is portable. D: self-serve terms of service are uniform, so the question is what they already permit. E: bespoke enterprise paper, where the change-of-control exposure alone can move the price.

P3 · Voice-of-customer and churn synthesis

What it is. Transcribed customer calls, churned-customer calls, support logs, reviews and win-loss notes synthesized into churn reasons, pricing headroom, product gaps and competitive position, with the calls themselves run at a volume the deal would not previously have afforded.

The line it moves. Price, and the assumptions in the revenue plan: gross retention, price realization and expansion.

Typical range and the evidence. Medium. The clearest number is a cost comparison from one vendor: traditional commercial diligence from the large consultancies at $500,000 to $1M, against $50,000 for an AI-voice-agent process, reported by TechCrunch with named founders and unnamed clients.¹⁰ The consultancy figure reads like large-cap pricing and the comparison flatters, but the point survives at any ratio: the price drop changes when diligence happens, so that a buyer can commission it "well before they have high conviction in a deal" rather than after the letter of intent. Third Bridge, an expert network with every incentive to say otherwise, puts manual transcript review at 10 to 20 hours per deal phase and argues that AI does not reduce the number of expert calls, it makes them "more focused and higher yield"; the saving is in synthesis, not in primary research.¹¹ Software Equity Group's diligence timeline gives customer calls about one week of a five-to-seven-week window between letter of intent and close, so this lever compresses cost far more than it compresses the calendar.¹²

Conditions. Access to the target's support and call data under the NDA, or a customer reference list large enough to sample; a synthesis prompt that asks for reasons and quotes rather than sentiment scores; a person who reads the transcripts behind any finding that moves the price.

Kill criteria. If the synthesized churn reasons do not match what the top ten churned accounts say when called by a person, the synthesis is set aside.

Horizon and stage. Deploy. Diligence.

How it translates. A: agencies name the scheduler they would have to hire without the product, which is the price anchor. B: customers describe the audit they passed, which is the retention anchor. C: merchants describe cash-flow pain, which is the payments anchor. D: self-serve users rarely take calls; the corpus is support tickets and reviews. E: enterprise references describe the implementation, which is the services anchor.

P4 · AI-exposure assessment

What it is. A structured judgment on whether the target is substitutable: how much of its revenue is seat-based and in roles that agents compress; where it sits in the customer's workflow; whether its moat is data, distribution, regulation or a user interface a model could replace; and what a competitor with a frontier model and no legacy could build against it.

The line it moves. Price, and whether to bid at all.

Typical range and the evidence. Medium, and the direction is not what the marketing says. Meritech's April 2026 index put the median public software company at 3.2x ARR, 63% below the pre-ZIRP median, with only five companies above 10x forward revenue and 80% below 5x, and found that the bucket of companies with AI tailwinds fell harder than the rest.¹³ Candriam's February 2026 note dates the compression from August 2025 and offers the resilient categories, cybersecurity, embedded enterprise platforms, regulated verticals and data infrastructure, without a single number; use it for the segmentation and Meritech for the figures.¹⁴ Vista's chief executive told Bloomberg in February 2026 that "the vast majority" of its companies were not experiencing meaningful churn to AI alternatives, a statement made by an interested party during a selloff, with no churn data and about billion-dollar-revenue businesses rather than the ones this book concerns.¹⁵ The Thoma Bravo admission cited on the hub is the other side of the same coin. RSM's diligence framework is the most usable structure I found: data quality, infrastructure, staffed talent, governance, cultural openness, and use cases with "a measurable line of sight to EBITDA in six months or less."¹⁶

The honest state of this lever is that the exposure is real, the frameworks are qualitative, and no threshold exists for the seat-based share above which a target should be repriced. The assessment is a judgment, made explicit and written down so that it can be wrong in a way that is later visible.

Conditions. The contract database from P2 (which revenue is per-seat, in which roles); usage data by role; a competitive scan that includes AI-native entrants; a stated view from the deal team on which of the target's functions a customer could replace with a model and a spreadsheet.

Kill criteria. None; this lever cannot fail, only be omitted.

Horizon and stage. Deploy. Diligence.

How it translates. A: regulated, workflow-embedded, seats are caregivers and schedulers; exposure low, moat regulatory. B: content moat that AI cheapens to build, so exposure is in the moat itself. C: exposure low if payments are attached, higher if not. D: the highest exposure on the map; seats in knowledge-worker roles, self-serve, replaceable UI. E: exposure concentrated in the services line, which AI compresses whether or not the owner wants it to.

P5 · Lever underwriting

What it is. The approved levers priced into the model before close, each with its baseline from the target's own data, a range, a kill date and its share of the foundation cost, and separated into base case and upside. It converts the value creation plan from a document written after the deal into an input to the price and the structure, including earn-outs or seller paper tied to lever outcomes.

The line it moves. Price and structure.

Typical range and the evidence. Strong for the fact that firms do it, weak for whether it works. The adoption surveys above say a majority of dealmakers use AI somewhere in valuation. Bain and StepStone's 2026 survey found diligence and sourcing to be the most frequently cited highest-return uses of the technology inside the firm, while 39% of GPs expected no material portfolio benefit in 2026.¹⁷ Bain's 2026 global report gives the arithmetic that makes underwriting the levers necessary rather than optional: "Typical deals now require around a 10% to 12% average annual growth in EBITDA to generate the same benchmark 2.5X return over five years," against about 5% in the 2010s, with holding periods "at around seven years."¹⁸ Grant Thornton's 2026 survey of a thousand senior leaders found 78% lacking confidence they could pass an independent AI governance audit within ninety days, a defensible base rate for what the buyer will find.¹⁹ A 2021 West Monroe survey, outside the window and cited only as a pre-AI baseline, found fewer than half of PE firms incorporating diligence outputs into the value creation plan at all.²⁰

Conditions. P1 through P4 done; a model that can carry lever-level lines rather than a single "AI savings" row; an operating partner who will own the base case after close.

Kill criteria. A lever that cannot be given a baseline from the target's data before close goes into upside, not the base case.

Horizon and stage. Deploy. Diligence.

How it translates. The same for all five: the four or five carrying levers from Part V go into the base case only if they have baselines. Everything else is upside.

P6 · Financial and quality-of-earnings acceleration

What it is. AI-assisted quality of earnings, revenue recognition review and cohort rebuild from raw ledgers, billing exports and contracts, with the accountant reviewing exceptions rather than building the schedules.

The line it moves. The cost and calendar of financial diligence.

Typical range and the evidence. Weak, and stated as such. The Big Four QoE benchmark above is the only hard number.³ No independent study compares AI-assisted and traditional QoE on the same target, and the accounting firms have published nothing beyond announcements. The mechanism is the same as O6 (close automation), where the evidence is strong, transposed to a one-time exercise, and the reason to include it is that on a small deal the QoE fee is a material share of diligence spend. The card is kept so that an owner who tries it measures it.

Conditions. Ledger and billing exports in usable form; a provider willing to disclose what the tool did and what the person did.

Kill criteria. Any reconciliation difference between the automated cohort build and the target's own reported ARR above 1% halts reliance on the automated output.

Horizon and stage. Deploy. Diligence.

How it translates. Material only for A and C, where the fee is large relative to the deal; a rounding error for D and E.

P7 · The hundred-day foundation

What it is. The handful of things every operating lever depends on, built between close and day one hundred and funded as one line: a warehouse joining billing, CRM, usage, support, finance and HR data at the customer and employee level; an event stream instrumenting product usage at the action level; a knowledge base maintained for agents; a model gateway that logs every prompt and output, enforces which models may see which data classes, tracks cost by agent and routes only to providers under the right agreements; an evaluation harness; identity and least-privilege permissions for agents; the agent register; and baselines for every metric in the plan, taken from the company's own data.

The line it moves. One-time operating expense, and the feasibility of every R, G, O, C and X lever.

Typical range and the evidence. Strong for the mechanism, because the components are standards rather than products. Anthropic's engineering guidance draws the line between workflows ("LLMs and tools are orchestrated through predefined code paths") and agents ("LLMs dynamically direct their own processes and tool usage") and advises "the simplest solution possible, and only increasing complexity when needed"; most of what year one deploys is workflow by that definition.²¹ OWASP names the failure the permissions design prevents: "Excessive Agency is the vulnerability that enables damaging actions to be performed in response to unexpected, ambiguous or manipulated outputs from an LLM," with the remedy to "limit the permissions that LLM extensions are granted to other systems to the minimum necessary."²² The evaluation harness follows Hamel Husain's three levels, assertions, trace review and A/B tests on the outcome the product is meant to move.²³ The legal basis is where the foundation is most often built wrong: as of July 2026 Anthropic's documentation states that "HIPAA readiness and ZDR cannot coexist on a single 1P API organization" and that its covered models "require 30-day data retention," and OpenAI offers a business associate agreement on its API and enterprise products but not its Business plan.²⁴ The retention term in the plan is the one the agreement actually offers, written into the register.

In the case study the foundation cost about $250,000 of one-time data and migration work plus platform subscriptions of roughly $350,000 a year at maturity and a three-person transformation office, all inside the R&D and G&A lines. Scale those with the company.

Conditions. Funded before close (P5); a transformation lead hired or seconded by day fifteen; the CEO's calendar.

Kill criteria. At the day-one-hundred review every initiative shows a baseline and a first measurement, and three numbers must have moved. If the warehouse cannot produce a customer-level view joining billing and usage by day sixty, the operating levers are re-sequenced behind it rather than launched on spreadsheets.

Horizon and stage. Deploy. Close to day one hundred.

How it translates. A: the gateway is a business associate boundary and the retention term is a compliance fact. B: the warehouse joins regulatory submissions to customers, and the register is customer-facing evidence. C: the event stream must carry transaction data, and PCI scope shapes the gateway. D: the foundation mostly exists; the register and the cost tracking by agent do not. E: multiple systems from multiple acquisitions make the warehouse the largest single item.

P8 · Pricing reset at close

What it is. The price and packaging change decided in diligence (from P3's headroom evidence and P4's pricing-basis split) and executed in the first renewal cycle after close, before any automation touches a customer-facing function. It is the first condition of the playbook made concrete: price before automate.

The line it moves. Revenue, at roughly 90% flow-through because delivery cost does not change.

Typical range and the evidence. Strong, because the largest vendors publish their increases. Salesforce announced that "on August 1, 2025, list prices will increase by an average of 6%" on its core clouds, alongside Agentforce add-ons "starting at $125 per user per month," and Slack's Business+ plan went "to $15 from $12.50 per user per month" the same week.²⁵ Zylo's 2026 index of more than 40 million licenses found the average organization's SaaS spend rose nearly 8% in 2025 while the application count held steady.²⁶ Growth Unhinged's May 2026 survey of more than 230 companies found that three in four changed pricing or packaging within the previous year and that hybrid pricing had risen from 25% to 37%.²⁷ Against that, the same survey found 70% of AI spend coming from existing technology budgets, which says the customer's wallet is not growing; price moves inside it.

The reset at close is distinct from R1, which is the multi-year program. The reset is the decision, and the reason it sits in this chapter is that it must be made before the automation that would otherwise make it impossible.

Conditions. The renewal calendar and contract terms from P2; the headroom evidence from P3; a control cohort so the retention effect is measured; a first visible new value (even a roadmap) to attach it to.

Kill criteria. Churn in the uplifted cohort more than one point above the control cohort pauses the program.

Horizon and stage. Deploy. Close to day one hundred, executed through year one.

How it translates. A: per-user pricing unchanged since 2021 against a product that runs the agency; the largest single lever. B: tiering with the premium tier carrying the regulatory content updates. C: subscription reset is secondary to payments attach, but the two are packaged together. D: the reset is the move from seats toward hybrid, and it is the riskiest of the five. E: services repriced into subscription, which is G3 executed as a pricing decision.

What goes in the model

A pre-close model that carries these levers has four features a conventional one lacks. It has a foundation line (P7) in year-one operating expense. It has the carrying levers from Part V as separate rows with baselines, ranges and kill dates, and the rest in a labelled upside case. It has the revenue plan's retention and price assumptions tied to P3's evidence rather than to a benchmark. And it has the AI-exposure judgment (P4) written down, with the seat-based share and the moat classification, so that the board can see in year two whether it was right.

The calendar does not compress as much as the cost. Diligence workstreams run concurrently, and Software Equity Group's five-to-seven-week window between letter of intent and close is set by legal negotiation and confirmatory work as much as by analysis.¹² The gain is that more of the work gets done inside that window, on more of the company, for less.